Platform v5.0 release notes
This page lists the customer-facing changes in each Platform v5.0 release. For an overview of the release, see What's new.
5.0.0​
Released November 18, 2026.
Platform v5.0 is a limited availability release, recommended only for new deployments. Keep existing deployments on Platform v4.12 and vCluster v0.37, the current LTS releases. For upgrade guidance, contact vCluster Labs.
The removed features, deprecations, and breaking changes below describe how v5.0 differs from v4.12.
Removed features​
- Shared secrets. Shared secrets, shown as Global Secrets in the UI, are
removed. Project secrets can no longer sync from a shared secret. A project
secret that synced from one keeps its last data and becomes an ordinary
project secret. The
imagePullSecretsfield on users and teams is removed too. Use project secrets instead, optionally synced from HashiCorp Vault. Vault is now configured per project only. Avaultkey left in the platform configuration is ignored, and the platform logs a message saying so. To keep managing shared secrets on v4.12, use a vCluster CLI compatible with v4.12. See the vCluster v0.38 release notes for the CLI changes. - Cost control. Cost control and its ROI dashboard are removed. A
costControlkey in the platform configuration is ignored. On startup, the platform uninstalls the Helm releases it installed for cost control:prometheusandopencoston each connected cluster, andglobal-prometheuson the cluster where the platform runs. Releases with those names that you installed yourself are kept. If other tools read metrics from these releases, set up your own monitoring before you upgrade. - Argo CD project integration. The project-level Argo CD integration, set
with
spec.argoCDon a project, is removed. The platform no longer creates an Argo CD project for each platform project. Use the cluster's Argo CD integration invcluster.yamlinstead. See Connect Argo CD. - Bundled apps. Installations no longer get the ArgoCD, jsPolicy, Ingress Nginx, Cert Manager, Prometheus Stack, and OpenTelemetry Collector apps. Apps themselves aren't removed, and the example Helm, Bash, and Manifests apps are still available.
- Moving clusters between projects. The
migratevirtualclusterinstancesubresource on projects, deprecated in v4.11, is removed.
Deprecations​
- Project namespaces (spaces) are deprecated. These are the namespaces a project manages through the platform, not Kubernetes namespaces in general. Platform management of them is planned for removal in a later v5.0 patch release. In v5.0.0 they keep working in projects outside tenants, and the UI shows a deprecation banner on the namespace pages. Tenant projects can't create or import them. After removal, existing namespaces and their workloads stay on their clusters, but the platform no longer manages them. Use clusters for workloads that need platform management.
- The Istio integration and sleep mode for Istio ingress are deprecated. Nothing is removed in v5.0.
- Metal3
bareMetalHosts.selectoris deprecated in favor ofmachines.selector. You can't set both. - The Terraform template variables
var.vcluster.requirementsandvar.vcluster.nodeEnvironmentare deprecated. Usevar.vcluster.propertiesandvar.vcluster.networkEnvironment. - The
DatabaseConnectorAPI is deprecated in favor of the Connector API. It still lists shared-database connectors.
Breaking changes​
Each item below changes an API, a configuration, or a behavior that v4.12 supports. Check manifests and automation written for v4.12 against them.
- A wildcard project member (
User/*orTeam/*) binds the project's scope group,loft:platformorloft:tenant:<tenant>, instead ofloft:authenticatedorloft:team:*. On a platform without tenants,User/*keeps its effect, butTeam/*now also reaches users who belong to no team. Agents that Platform doesn't upgrade refuse wildcard members on direct endpoints until upgraded. - SSH keys require
metadata.generateNamewhen created. A fixedmetadata.nameis refused, including in declarative manifests. After creation,spec.publicKey,spec.owner, andspec.accessare immutable. Create a new key to replace them and update references to its generated name. - The
cluster-collectorArgo CD application template, shown as Fleet Observability: Cluster Collector in the UI, reads credentials from themetrics-writerSecret. ItsotlpEndpointandbearerTokenparameters are deprecated. An application that sets only one of them fails to render withBuildApplicationFailed. See Configure edge collectors. NodeEnvironmentis replaced byNetworkEnvironment. The Terraform node provider fieldnodeEnvironmentTemplatebecomesnetworkEnvironmentTemplate.infrastructure, and administrators create network environments instead of the platform creating one per cluster. The old template'skubernetessection has no replacement.- The default Metal3 chart moves from 0.4.1 to 0.7.3. Ironic's database moves
to a MariaDB StatefulSet by default, and
mariadb.modemust bestatefulsetorexternal. Bundled Metal3 no longer usesironic.ipaAdvertiseIPor a LoadBalancer for IPA, and Neutron needsneutron.networks.provisioning. Servers reach the IPA API through the DHCP proxy. See Metal3 requirements. - Metal3 no longer filters hosts by inspected CPU or memory against the node
type's
resources. Node types that relied on that filter can now select smaller hosts. Use labels and selectors to express hardware size requirements. Theresourcesfield still informs scheduling and cost. - The Platform API Go module moves from
github.com/loft-sh/api/v4togithub.com/loft-sh/api/v5. - Stack templates use
parameters. A template written for a v4.12 pre-release that still usesinputsis refused with an unknown field error. Platform v4.12.0 and later already useparameters. - Node claims can't set the properties their node provider controls. Metal3
claims can't set the network CIDR, IP range, or DNS server properties. NICo
claims can't set the OS image ID, and KubeVirt claims accept only an allowed set. The API
refuses such a claim, and an Auto Nodes claim fails with
PropertiesNotAllowed. A claim'sccm-enabledproperty is ignored, and the node provider's setting applies. - The
vcluster.com/user-data-template-secretandvcluster.com/network-data-template-secretproperties are removed. Put the data in a MachineConfigTemplate and use the matching-configproperty. See Machine config templates.
New features​
Tenants​
- Tenants. A Tenant is a customer-organization boundary above projects. A
platform admin grants each Tenant the node types, OS images, templates, SSH
keys, connectors, and control plane clusters it may use, and lends machines
and network environments to it. Node types, OS images, control plane
clusters, and connectors are granted by name with
allow.byNameor by label withallow.byLabels, but not both. Templates are granted by label. See What are tenants. - Tenant administration. Each Tenant gets an admin user, an admin team, and a default project. Tenant admins manage their own users, teams, and projects, and roles added inside a Tenant must be management role templates. Tenant admins can impersonate users within their own tenant. See Create and configure a tenant.
- Tenant access to control plane clusters. A tenant user reaches a control plane cluster through the platform proxy only when the cluster is assigned to their tenant and a cluster access admits them. See Cluster access.
- Per-tenant branding. Each Tenant can have its own logo and colors in the platform UI. See Brand a tenant.
- External cluster imports per tenant. A platform admin decides per Tenant
whether its users can import clusters deployed outside the platform, with
spec.externalClusters.enabled. See Allow importing externally deployed clusters. - Tenant deployment boundaries. Inside a Tenant, apps, stacks, and Argo CD
applications can't deploy onto a control plane cluster or with target
host. A Tenant's KubeVirt VMs run in a namespace named<tenant>--<namespace>, so Tenant names can't contain--. See Tenant boundary reference.
Machines and bare metal​
- Node pools. A NodePool defines the nodes for a cluster with private nodes
in the platform instead of in
vcluster.yaml. Cluster creation, templates, and the cluster's Node Pools tab manage them. Node pools need the Auto Nodes license feature, plus Machine Management when the node provider draws from the machine inventory. Node pools need vCluster v0.38 in the cluster. See Node pools. - Machine allocation policy. Metal3 and NICo node types choose between
Assignedplacement on machines reserved for a tenant andOnDemandplacement on unassigned machines only. Metal3 and NICo node types default toAssigned, and KubeVirt node types default toOnDemand. A node type's own policy wins, then the node provider'sspec.machineAllocationPolicy. Metal3 and NICo placements on machines assigned to the Tenant don't count toward its node type quota, and KubeVirt placements always count. See Machine types and node types. - Machine selection by label. Metal3 node types select Machines with
machines.selector, and BareMetalHost labels sync to Machines. See Machines. - NetBox inventory sync. A NetBox connector imports tagged NetBox devices into the machine inventory for the Metal3 provider. See Connect to NetBox.
- OS image store. Upload OS images to a store and provision machines from them. See OS images.
- Redfish Explorer. Platform admins can browse a host's Redfish resources and run BMC actions from the platform UI. See Inspect the BMC.
- Power actions for bare metal machines. Metal3 and external platform
machines can be started, stopped, and hard rebooted, as KubeVirt VMs already
could. Use Start, Stop, and Reboot in the UI, or set a node
claim's
spec.power.stateor itsmachines.vcluster.com/rebootannotation. - Network environments. A NetworkEnvironment describes the network that machines join. Administrators create them and lend them to Tenants. See Network environments.
- Aviz network integration. The Metal3 provider can allocate servers to an
Aviz ONES tenant through a network environment. It needs the
avizlicense feature, Neutron mode, and an explicit network CIDR. Network segmentation depends on your ONES fabric configuration. See Metal3 network integration. - Node claim timeout. A node claim waiting for a machine during provisioning
fails after 30 minutes by default. Set
machines.vcluster.com/provisioning-wait-timeoutto change it. Capacity waits during scheduling retry without a time limit. - Node claim diagnostics. Node claims report why provisioning is stuck or failing. See Machine lifecycle.
- Host devices and GPU passthrough for KubeVirt VMs through the vCluster Device Operator. Platform v4.12 deployed operator chart 0.0.7, and v5.0 deploys 0.2.2. See Host devices.
Stacks​
- Stacks. Platform v5.0 completes the Stack workflows for administrators and tenants, including RBAC, templates, and lifecycle. See What are stacks.
- Certified stacks. The certified bundle adds NVIDIA NVKE Management Plane.
The bundle comes from the
loft-sh/vcluster-stacksrepository and includes NVIDIA Dynamo and Saturn Cloud Enterprise, also in v4.12.5. See Certified stacks.
Connectors and observability​
- Connector API. Connectors are now managed through a new Connector API, and platform admins can grant connectors to Tenants. Two connector types are new in v5.0: NetBox, for machine inventory, and OS image store, for uploaded OS images. They join the shared database, Argo CD, and observability connectors. See Connectors.
- Telemetry destinations. The observability gateway forwards telemetry to additional destinations, and edge collectors can collect pod logs. See Install the observability gateway.
Operations​
- Management backup keeps identities. A platform restored from a management backup keeps its identity. Also in v4.12.4. See Back up the platform.
- Administrator recovery. Recover access when no platform admin can sign in. See Reset the admin password. Also in v4.12.2.
- Argo CD cluster labels. Clusters registered in Argo CD carry project and
instance labels. Add your own with
spec.argoCD.cluster.metadataon a control plane cluster, or withintegrations.argoCD.cluster.metadatain a cluster'svcluster.yaml. This makes ApplicationSet cluster generators usable. See Connect Argo CD. Also in v4.12.2.
Changes​
- Metal3 skips OS image TLS verification by default. Set
ipaImages.insecure: falseto verify. - Users who can't update node providers and node types see a reduced, read-only view of them.
- Automatic snapshots retry through connection errors instead of failing. Retention also deletes the archive of a pruned failed snapshot from storage. See Retries after connection errors.
- A cluster now fails, instead of staying pending, when its database connector fails.
- The platform refuses to start when
LICENSE_TOKENis set but empty, for example from a Secret that isn't populated yet. The pod restarts until the Secret has a value. Before, the platform fell back to activation, a cached license, orLICENSE_KEY. helm registry loginfor OCI registries now verifies TLS, unless the chart setsinsecureSkipTlsVerify. Before, the platform skipped verification on login. A registry with a certificate the platform doesn't trust now fails login until you add its CA or setinsecureSkipTlsVerify.
Fixes​
- Deleted clusters no longer count toward per-user quotas. Also in v4.12.4.
- Named SSO connectors honor
auth.disableUserCreation. Also in v4.12.6.
Known issues​
- An OS image upload in the browser can't resume after the page reloads, and an upload can't be canceled.
- Deleting an OS image waits while its store is unreachable.
- The observability gateway keeps queued telemetry in memory, so a restart loses it.